Issue with ReFS on Windows Server 2025

Hi Francois Gotti,

How was your issue going? Have you been able to solved it yet? Is there anything I can help you with? Please let me know.

VP

I have checked that, but still in monitoring because with number of users connected this last week end. Impossible to see if better. I will check during the week.

Since you are waiting for the production load to hit, I think I should give you a specific "Early Warning System" to set up now. Do not wait for users to call you saying "it's frozen."

1. The "Early Warning" Counters (Set this up in Performance Monitor)

If the server is going to freeze, these specific counters will spike before the user experience degrades. Open perfmon.exe and add these:

• ReFS\Metadata Scavenger: Watch the % Used or Queue Length.

ReFS on Server 2025 is much more aggressive about "cleaning up" metadata than 2022. If this queue grows without shrinking, the driver is choking, and a freeze is imminent.

• Logical Disk\Current Disk Queue Length: Threshold: If this stays consistently above 5 (assuming an SSD array), your storage controller isn't acknowledging writes fast enough for the new ReFS driver.
• Memory\Pool Non-Paged Bytes: ReFS consumes Non-Paged Pool memory for metadata mapping. If you see a "sawtooth" pattern (rapid rise, sudden drop) that keeps getting higher, you have a memory leak in the filter driver.

2. Why Server 2025 is worse than 2022? Server 2022 typically runs ReFS v3.7. This version was "lazy" with metadata. It would cache a lot in RAM and write it to disk later. This felt faster but was riskier for data loss during power cuts. Server 2025 introduces ReFS v3.12 (or higher). This version forces "Integrity Stream" checks and metadata flushing much more frequently to prevent corruption. Citrix Profile Management (CPM) performs massive amounts of tiny metadata operations (opening files, checking timestamps). ReFS v3.12 tries to commit every single one of these to disk immediately, overwhelming the queue.
3. Since you switched to Citrix Profile Management (CPM), there is a specific feature that is famous for killing ReFS performance, which FSLogix did not have enabled by default.

Please verify this in your Citrix Policies immediately:

1. VHDX Disk Compaction:

Setting: "Reclaim free space on VHDX containers."

Action: DISABLE this during production hours.

This forces the server to mount the VHDX, defragment it, and zero-out blocks while the user is logged off but the disk is still hot. On Server 2025 ReFS, this background IOPS storm can freeze the whole volume.

2. Active Write Back: If you have this enabled, it bypasses the localized cache and writes directly to the share. Ensure this is Disabled so that writes are buffered.

My recommendation for this week is to keep perfmon open. If you see the Metadata Scavenger queue spike, you may need to apply a specific registry key to force ReFS back to "Lazy" mode (specifically RefsDisableLastAccessUpdate = 1), but let's test your current changes first.

Keep me informed, thanks! Have a nice day!

VP